Proceedings of the 2008 IEEE, CIBEC'08 978-1-4244-2695-9/08/$25.00 ©2008 IEEE

VLSI SYNTHESIS OF HETEROGENEOUS AND SIRM FUZZY SYSTEM FOR CLASSIFICATION OF DIABETIC EPILEPSY RISK LEVELS

R. Harikumar1 , A. Shanmugam2 , Parthiban Rajan 3

1Department of ECE, Bannari Amman Institute of Technology, Sathyamangalam. 2The Principal, Bannari Amman Institute of Technology, Sathyamangalam.

3UG scholar, Bannari Amman Institute of Technology, Sathyamangalam.

Abstract-The objective of this paper is to design, simulate, and synthesis a simple, suitable and reliable VLSI fuzzy processor for classification of diabetic epilepsy risk levels. The performance of three different fuzzy techniques are analyzed and compared. While designing the fuzzy processor the Cerebral Blood Flow (CBF), EEG signal features and aggregation operators are taken as parameters. The classification of risk level is based on clinical data and observation. Three different fuzzy techniques with minimum rules such as a two input heterogeneous fuzzy technique, Single Input Rule Models (SIRM) are analyzed. The parallel architecture is incorporated in this design with independent functional units. These functional units process the data simultaneously by which the processing speed is enhanced. The SIRM fuzzy system with Bell input – Bell output, and Bell input –Triangle output are simulated and synthesized for various values of Cerebral Blood Flow using VHDL. The simulated and synthesized Field programmable Gated Array (FPGA) fuzzy processor closely follows the mat lab version. Keywords -Fuzzy Processor, VLSI, FPGA synthesis, SIRM system, Epilepsy risk levels

I. INTRODUCTION

Fuzzy models have supplanted many conventional technologies in various scientific applications and engineering systems, especially in control systems and pattern recognition. The approach to implement fuzzy logic systems may be software only, hardware only, or a mixture of both. Suitability of an implementation approach depends mostly on the application type and performance requirements [1]. The partitioning of the solution to its software and hardware parts can be done with respect to different criteria for example, response time, reliability, price etc. The software part requires a processor, on which it will run, while the hardware part can be in the form of functions implemented as functional units and used as the standard blocks available at design time from a digital design library. In recent years, Field Programmable Gated Array (FPGA) technology has been used to implement fuzzy logic for solving real world problems, such as image processing, fuzzy database, medical diagnosis, and Industrial engineering applications.

The goal of this paper is to design, simulate, and synthesis a simple and robust VLSI fuzzy processor to classify the epilepsy risk levels of diabetic patients. The Cerebral Blood Flow (CBF) measurements, Electroencephalogram (EEG) signals are the input

Parameters. The slope of fuzzy membership function is decided by the aggregation operators. Sudden, recurrent and

transient disturbances of brain functions or movements of body that results from excessive discharging of groups of brain cells characterize epilepsy. In clinical neurological practice, detection of abnormal EEG activity plays an important role in diagnosis of epilepsy [2]. Quantitative EEG studies are used in cerebro-vascular disorders to improve diagnostic sensitivity. Based on the well-established theory of CBF mechanism, as CBF increases to 70 percent, the oxygen concentration on the blood stream is reduced by 30 percent [3]. This leads to long lasting hypoxia and cerebral ischemia. Therefore, the epileptic convulsion risk is increased in diabetic patients [4]. EEG can easily detect the oxygen delivery and utilization in brain. This indicates the correlation between CBF with EEG signals.

In Fuzzy St Theory (FST) this problem takes the form as below

Maximize FµΩ (x)= Agg µA1(x1), µA2(x2)….. µAn(xn)……(1) ` Where Agg stands for an appropriate aggregation operator. It combines the membership values in the set A1, A2,….An into the membership value of the set Ω formed by some operations on the sets A1, A2,….An. All aggregation operators are equivalent to the distance of the ideal one or anti ideal zero in the relevant metric. The organization of the paper is as follows; the section 1 introduces the concept for Fuzzy system design. The materials and methods are explained in the section 2. VLSI Design and simulation of heterogeneous Fuzzy system is discussed in section 3. Section 4 elucidates the FPGA Implementation of Fuzzy system. Results are discussed in the section 5and the paper is concluded in section 6.

II. MATERIALS AND METHODOLOGY

The basic block diagram of the fuzzy epilepsy risk level classifier is shown in Fig.1.The measured CBF will have five linguistic levels such as very low, low, medium, high and very high [5]. This is the input of the fuzzy system. Features such as energy, peaks, spikes and sharp waves and events are extracted from EEG signals, and are classified as more relevant variables Y. The clinical parameters such as index of convulsions, seizure timings and total body fatigue are less relevant variables set X. This confidence level of derived CBF will act as near equivalent to membership function of derived epilepsy CBF level which is another input to the fuzzy system.

Proceedings of the 2008 IEEE, CIBEC'08 978-1-4244-2695-9/08/$25.00 ©2008 IEEE

Figure 1 The basic functional diagram of diabetic epilepsy

level classifier 2.1 Fuzzy System and Data Collection

The basic structure of fuzzy system includes four main modules [6] such as Fuzzifier, defuzzifier, inference engine and knowledge base. The CBF is measured by tetra polar method [7]. For an adult, the normal blood flow in the brain tissue averages 38-45ml blood/100gm brain per minute. In the entire brain it comes about 750-900 ml per minute or 15% of the total resting cardiac output [4]. Raw data of EEG signals level of 16 bipolar channels are recorded. The following features are extracted from EEG signal for each epoch.

Energy of the signal Ei 1

n

Si2

Where Si is the sample value in micro volts [3]. The total number of spikes and sharp waves in all channels together are recorded as events 2.2 Fuzzy Membership

The fuzzy (Mamdani type) system with two inputs and one output along with the Center of Gravity (COG) Defuzzification method is selected. The measured CBF input is fuzzified with five linguistic variables like very low, low, medium, high and very high using triangular membership functions [7]. The other input is derived epilepsy CBF of the same patient obtained from extracted features of EEG signals [8], whose membership functions are bell shaped one, with five linguistic variables. The slope of the bell shaped function is derived from seventeen simulated test conditions using aggregation operators. This fuzzy system performs well with the following five Fuzzy Rules (Fr) in the rule base such as

Fr1: IF CBF is very low AND epilepsy CBF is very low, THEN output epilepsy risk level is normal

Fr2: IF CBF is low AND epilepsy CBF is low, THEN output epilepsy risk level is low risk

Fr3: IF CBF is medium AND epilepsy CBF is medium, THEN output epilepsy risk level is medium risk

Fr4: IF CBF is high AND epilepsy CBF is high, THEN output epilepsy risk level is high risk

Fr5: IF CBF is very high AND epilepsy CBF is very high, THEN output epilepsy risk level is very high risk

The linguistic fuzzy membership functions are illustrated in figures 2a, 2b and 2c.

Figure 2a. Input membership function (measured CBF) of the heterogeneous fuzzy system

Figure 2b. Input membership function (derived epilepsy CBF) of the heterogeneous fuzzy system

Figure2c. Output membership function (epilepsy risk level) of the heterogeneous fuzzy system

III. VLSI IMPLEMENTATION

In the last two decades or so, by far the strongest growth area of the semiconductor industry has been in silicon VLSI technology. The sustained growth in VLSI technology is fueled by continued shrinking of transistors to ever-small. The benefits of miniaturization, higher packing densities, higher circuits speeds and lower power dissipation have been key in the evolutionary progress leading to today’s computer and communication systems that offer superior performance, dramatically reduced the cost per function and much reduced physical size in comparison with their predecessors.

3.1 Fuzzy Processor Architecture Figure 3, shows the Fuzzy processor’s logical architecture for VLSI simulation and mentioned by Giuseppe ascia et al [9]. The internal organization includes the following

blocks: Figure 3. Fuzzy Rule Architecture for VLSI Simulation

Fuzzy set base- A digital memory that contains the fuzzy sets related to input variables with codes stored in the rule memory. Its internal organization allows the fuzzy set base to directly supply the core with the membership degrees (alpha

Proceedings of the 2008 IEEE, CIBEC'08 978-1-4244-2695-9/08/$25.00 ©2008 IEEE

values) of the input variables during the computation of a premise for each rule. Likewise, during conclusion processing, the fuzzy set base supplies the relative fuzzy set. Core- it assesses a stream of fuzzy rules in the pipeline and possesses an internal organization that allows parallel assessment of up to four antecedents per rule. The core can also store partial processing results if more than one rule involves the same inference. Defuzzifier- This unit forms the defuzzification block and provides an output value proportional to the Maximum of fuzzy inference output. Control unit- Coordinator of the processor’s internal activities, this unit also interacts with components outside the processor through the signals mentioned earlier. Rule memory- For passing information about a rule to the processor, this memory contains the code of the membership functions associated with each variable in fuzzy rule’s premise and conclusion. The processor communicates through the

• Input bus, through which the processor receives, input values; rule memory data, a digital bus through which the processor receives information it needs to identify the membership function in an active rule’s premise and conclusion;

• A rule memory address, which stores a KBM-generated value that points to the rule memory location containing information about the active rule to be processed;

• The control bus, a set of digital lines through which it is possible to control fuzzy processor functions and supply the necessary signals for interactions outside the processor; and

• The output bus, an analog signal representing the defuzzified output value.

The VLSI design procedure of Fuzzy processor is discussed in the following section of the paper. 3.2 VLSI Design Procedure of Fuzzy Processor

The VLSI design and simulation using Verilog HDL (Hardware Description Language) [10], [11] is undertaken for the Heterogeneous fuzzy system for the classification of epilepsy risk level in diabetic patients. This fuzzy system [12] is modeled with input and output ports. The other necessary variables are declared as memory registers to hold the processing values. The bell shaped membership functions are evaluated from the functional parameters like deviation slope and center value of measured CBF is modeled [12]. To avoid the complexity and to know the working nature, primarily the system is developed for simulation by implementing Integer multiplication and division in which the parameters are multiplying by scaling factors to avoid floating values. As for as the simulation result is concerned the system performs well and gives correct results in all linguistic regions of the cerebral blood flow. Table 1 shows VHDL simulation results for the Two input Heterogeneous fuzzy system. Towards the better understanding of simulation process of the heterogeneous fuzzy system a sample of six CBF values in the different regions are selected and analyzed.

Table1. Simulation Results for Heterogeneous system

Table 1 depicts the calculation of membership function and the formation of Rule base for a given value of CBF. If the given CBF placed in the overlapping region then all the selected linguistic variable membership functions were calculated. The output risk level is calculated by MAX defuzzification process. For example if the measured CBF is 70 ml. This falls under two linguistic variables (Low and Medium) in Measured CBF. Hence, both the membership values are calculated, finally maximum membership value is extracted by KBM generated values. Likewise, membership function values are calculated for Derived CBF also. Maximum value is decided as an output epilepsy risk level value for the derived CBF. Finally output epilepsy risk level is determined by rules in the rules base. To Establish the best output risk level state in the overlapping region of the linguistic variables the above mentioned simulation is performed. 4. Application of Single Input Single Output (SISO) Fuzzy System Model. The conventional fuzzy inference model which puts all the input items into the antecedent part of each fuzzy rule, causes the total number of possible fuzzy rules to increase exponentially with the number of the input items and has difficulty in setting up each rule [12].The SIRM’s dynamically connected fuzzy inference model is applied, to over come these problems. This defines a SIRM separately for each input item as [1]

SIRM-i:Rij:if xi=Ai

j then fi = Ci jmij=1. --------- (2)

Here SIRM –i denotes the SIRM of the i th input items, and is the Ri

j ith rule in the SIRM –i. The ith input item xi is the only variable in the antecedent part, and the consequent variable fi is an intermediate variable corresponding to the output item f.Ai

j and Cij are the

membership functions of xi and fi in the j th rule of the SIRM-i. In this fuzzy model we applied the measured CBF as an input with bell shaped membership function whose slope function (1/γ) is the same one used in derived epilepsy CBF membership function which is associated with the degree of importance of the clinical and EEG signal parameters. As we stated above this

Proceedings of the 2008 IEEE, CIBEC'08 978-1-4244-2695-9/08/$25.00 ©2008 IEEE

SISO fuzzy system is closer to the dynamic SIRM fuzzy system only in the sense of degree of importance of the derived epilepsy CBF input parameters. SISO fuzzy system is different from conventional SISO fuzzy system and it is closer to the dynamically connected SIRM fuzzy system. Therefore we call this a SIRM fuzzy system and the same is depicted in figure 5.

Figure 5. SIRM Fuzzy System Model

The applied SIRM fuzzy system is designed with five rules in the rule base and measured CBF as the only input.

SFR 1) IF CBF is very low THEN output epilepsy risk is normal

SFR 2) IF CBF is low THEN output epilepsy risk is low risk

SFR 3) IF CBF is medium THEN output epilepsy risk is medium risk

SFR 4) IF CBF is high THEN output epilepsy risk is high risk

SFR 5) IF CBF is very high THEN output epilepsy risk is very high risk

4.1 Membership Function of SIRM Fuzzy System The membership function of measured CBF is bell shaped and slopes function (1/γ) is embedded with EEG signal information .The output epilepsy risk level of this system uses triangular membership function. The input membership function of the SIRM fuzzy system is shown in figure 6.

Figure 6. Input membership function of sirm fuzzy system

4.2 Simulation of SIRM fuzzy system The simulation result of the SIR (Bell input -Triangle output) fuzzy system is shown in Table 2.

Table 2. Simulation Results - SIRM Bell Input Triangle Output Fuzzy System

This SIRM system is analyzed for various measured CBF values. By using the RULEBASE output triangle membership state is defined. Table 2 shows the example for the SIRM Bell input- Triangle output fuzzy system. If the input CBF is 83.4 ml. It falls under the two linguistic sets (Low and Medium) of Bell membership function. Therefore both the sets membership functions are calculated and maximum membership value of linguistic set is extracted as a Output epilepsy risk level.. This procedure is repeated for different CBF values .he simulation results of the SIRM-Bell input Bell output fuzzy system is shown in Table 3

Table 3. Simulation Results - SIRM Bell input Bell output fuzzy

system

SIRM Bell Input -Bell output fuzzy system is also tested for different values of CBF. To justify the output Bell state Max method is used from the calculated membership function values. 4.3 VHDL Synthesis and Programming Xilinx board (SPARTAN 3) Using ISE The synthesis process transforms the VHDL model into a gate –level net list. The target net list is assumed to be a technology – independent representation of the modeled logic. The target technology contains technology – independent generic blocks such as logic gates and register-transfer level (RTL) blocks, such as arithmetic-logic-units and multiplexers, comparators interconnected by wires. In such a case, a second program called RTL module builder is necessary. The purpose of this builder is to build, or acquire from a library of predefined components, each of the required RTL blocks in the user-specified target technology. Having produced a gate-level net list, a logic optimizer reads in this net list and optimizes the circuit for the user-specified area and timing constraints. These area and timing constraints may also be used by the module builder for appropriate selection or generation of RTL blocks. Different synthesis system support different VHDL subsets for synthesis. Since there is no direct object in VHDL that means a latch or a flip-flop, each synthesis system may provide different mechanism to model a flip-flop or a latch. Each synthesis system defines its own subset of VHDL language including its own personalized modeling style.

Spartan -3 family offers densities ranging from 50,000 to five million system gates. It is programmed by loading configuration data into robust, reprogrammable, static CMOS configuration latches (CCL) that collectively control all functional elements and routing resources. Spartan-3 FPGA platform also allows the user to make significant changes while keeping original device pin outs thus

Proceedings of the 2008 IEEE, CIBEC'08 978-1-4244-2695-9/08/$25.00 ©2008 IEEE

eliminating the need to re-tool PC boards. We can easily upgrade, modify, and test the designs even in the field itself. Embedded capabilities make Spartan-3 devices ideal as coprocessors or pre-and post-processors, offloading highly computational functions from a programmable DSP to enhance system performance. 4.4 Synthesis of Fuzzy System

The synthesis part of the fuzzy system is depicted in the figure 9 which includes 3 main blocks, Triangle membership function, Bell Membership function, and Fuzzy state level comparison. The summary of Mapped resource for the Heterogeneous fuzzy system synthesis is tabulated in Table.4 and shows that only less part of the resources is used for the VHDL synthesis process.

Table.4 MAPPED RESOURCES SUMMARY

Figure 9 shows the Entity diagram for the Fuzzy system which uses three main blocks to evaluate (i) Triangular membership function, (ii) Bell Membership function, and (iii) output state selector. Here both the Triangle and Bell membership blocks are assigned with the common CBF and Clock signals. Figure 9 depicts only the major Entity blocks of the heterogeneous Fuzzy system. But each block internally uses numerous components to synthesis the output Function. By using the RTL schematic internal components also be analyzed. Figure 10 shows the Internal Diagram of Triangle membership function is inferred by 16 D-type flip-flop(s), 10 Comparator(s). It also inferred that 53 D-type flip-flop(s)and 15 Comparator(s) are used for the Bell Membership function. SIRM Bell input Bell output fuzzy system uses number of Flip flops, Comparator and gates for the synthesis Process. MACRO STATISTICS of Total no of components used for this Fuzzy system is tabulated in the table.5.

Figure 9 Synthesis of Fuzzy system

Figure 10 Internal diagram of Triangle Membership function

Table.5 MACRO STATISTICS of SIRM Fuzzy system

Figure 11 Internal sub diagram of SIRM (Bell Input -Triangle output) fuzzy system.

IV. RESULTS AND DISCUSSIONS

Heterogeneous and SIRM fuzzy system operation with Triangle and Bell shape input membership function are analyzed by various methods Like Matlab, VLSI simulation, and FPGA synthesis. All these approaches provide different types of results. Hence a Performance comparison is needed whereby the advantages of one over the other can be easily validated and the best method found out. A Graphical representation of performance of the MATLAB modeled heterogeneous Fuzzy system, VLSI simulation and FPGA synthesis is compared in “Fig .12.” From this comparison it inferred that FPGA synthesis also closely follows the existing MATLAB with higher degree of Performance. SIRM fuzzy system is effective for all risk level classification when compare to the two input heterogeneous fuzzy system which miserably fails in Low and High epilepsy risk level classification. Several reasons to justify SIRM method are

i. Higher rate of classification ii. Lower False Level

iii. System closely follows the MATLAB

Proceedings of the 2008 IEEE, CIBEC'08 978-1-4244-2695-9/08/$25.00 ©2008 IEEE

9293949596979899

100101

Normal LowRisk

Medium HighRisk VeryHigh RiskPe

rcen

tage

of P

erfo

rman

ce

Matlab

VLSI

FPGA

Figure 12 Performance of Heterogeneous Fuzzy System in MATLAB, VLSI Simulation and FPGA synthesis

6. CONCLUSION

VLSI design and synthesis of three types of fuzzy

systems such as two input heterogeneous, SIRM ( bell input- triangular output), and SIRM ( bell input- bell output) were under taken to classify the diabetic epilepsy risk levels. For different CBF values the above systems are tested. The results show that the average performance is 98.28 and which closely follows the performance of system developed by MATLAB simulation. The VLSI simulation is done with arithmetic operations for scaled up parameter values, which is actually floating point value. Synthesis of this design is difficult because of the scaled up values. This will increase the simulation time and the processing time in synthesized hardware. Hence the system is suitable for off-line diagnosis. For on-line diagnosis, the same design can be synthesized by incorporating floating point arithmetic operations which may settled with better results. SIRM (bell input – bell output ) fuzzy outperforms the other two types of fuzzy systems. Further research is in the direction of Application Specific IC (ASIC) implementation of SIRM(bell input- bell output) fuzzy system which effectively enhance the performance in the time complex city of the system

ACKNOWLEDGEMENT

The authors express their sincere thanks to the Management and the Principal of Bannari Amman Institute of Technology, Sathyamangalam for providing the necessary facilities for the completion of this paper. This project is also funded by All India council of technical education (AICTE), Govt. of India, New Delhi, under Research Promotion Scheme (RPS), REF.NO.:8023/BOR/RPS-76/2006-07, dated 28th Feb 2007.

REFERENCES

[1]. Zoran Salcic, High Speed Customizable Fuzzy logic Processor: Architecture and Implementation, IEEE Tran. System Man Cybernetics Vol.31, No. 6, pp 731-736, November 2001.

[2]. Clement C Pang et al, A Comparison of Algorithms for detection of spikes in the EEG, IEEE Trans. on Bio-Medical Engineering vol.50,no.4,pp 521-526, 2003.

[3]. Alison A.Dingle et al, A multi stage system to detect epileptic from activity in the EEG,IEEE Trans. on Bio-Medical Engineering,vol.40,no.12,pp 1260-1268,1993.

[4]. Muro et al, A mathematical model of cerebral blood flow chemical regulation - part II, IEEE Trans. on Bio-medical Engineering vol.36,no.2,pp 192-201,1989.

[5]. Pamela McCauley-Bell, Adedoji.B, Fuzzy modeling and analytical hierarchy Processing to quantify risk levels associated with occupational injuries, Part I, IEEE Trans. on Fuzzy Systems ,vol.4,no.4,pp 124-131,1996.

[6]. R.Harikumar, S.Selvan, Fuzzy based classification of patient state in diabetic neuropathy using cerebral blood flow, J. Systems Society of India- Paritantra, vol.7,no.1,pp 37-41,2002.

[7]. R.Harikumar, Dr. (Mrs.) R.Sukanesh, B.Sabarish Narayanan, Application of aggregation operators in fuzzy logic based classification of diabetic epilepsy risk level, Proc. of Annual Convention and Exhibition (ACE) IEEE India Council ACE 2003, Pune, 2003.

[8]. Ronald R.Yager, On ordered weighted averaging aggregation operators in multi criteria decision making, IEEE Trans. on Systems Man Cybernetics, vol.18,no.1,pp 183-190,1998.

[9]. Giuseppe ascia, Vincenzo catania, Marco Russo, Lorenzovita, Rule Driven VLSI fuzzy processor, IEEE Micro, Vol.16, no.3, pp 62-74, June 1996.

[10]. Mahamoud A Manzoul, D.Jayabharathi , FPGA for Fuzzy Controllers, IEEE Tran. System Man Cybernetics Vol.25, No. 1 Jan. 1995.

[11]. K.Paramasivam, R.Harikumar, K.Gunavathi Simulation of VLSI Design using Parallel Architecture for Epilepsy risk level Diagnosis in Diabetic Neuropathy, Proc. Of National conference on VLSI Design and Testing, Coimbatore, India, Feb.21st and 22nd 2003.

[12]. Dr.(Mrs.).R. Sukanesh, R.Harikumar Performance Analysis of Different Fuzzy Techniques in Classification of Epilepsy Risk Level for Diabetic Patients using Cerebral Blood Flow, Aggregation Operators and EEG Signals, I.E .India Journal of Interdisciplinary panels vol.87, no 2, pp 12-20, May 2007.

[13]. Leo P. Karall Md., Joslin Diabetes Manual, LEA & FEBIGER, Philadelphia London 1989, Chapter 16.

[14]. Arthur C. Guyton, Textbook of medical physiology, Prism Books Pvt. Ltd Bangalore, 9th Edition, 1996.

[15]. Javier.G.Marin.B, Qiang Shen, From Approximate to descriptive fuzzy classifiers, IEEE Trans. on Fuzzy Sytems,vol.10,no.4,pp 484-497,2002.

[16]. Hauqn, Jean Gotman, A patient specific algorithm for detection and onset in long term EEG monitoring - Possible use as a warning device, IEEE Trans. on Bio Medical Engineering, vol. 44,no.2,pp 115-122,1997.

[17]. Gleb Beliakov, Jim Warren, Appropriate choice of aggregation operators in fuzzy decision systems, IEEE Trans. on Fuzzy systems, vol. 9,no.6,pp 773-784 ,2001.

[18]. K.P.Adlassnig, Fuzzy set theory in medical diagnosis, IEEE Trans. on Systems Man Cybernetics,vol.16,no.3,pp 260-265,1986.